Innovative Serve Shot Analytics in Table Tennis Using CNN-LSTM Framework

This research study presents a new structure comprising CNN and LSTM networks for identifying serve strokes in table tennis. Analyzed in this study were serve shots, a crucial element of games, utilizing a dataset of 10,000 shots grouped into four types: The basic spin, which includes top spin and back spin, side spin and no spin. CNN-LSTM structure unites a potential for extracting the spatial characteristics of images of CNNs and the capability to process the temporal sequence of shots of LSTMs to analyze serve shots efficiently. The proposed model outclassed the existing classifiers with a significantly higher accuracy of 92%. In contrast, the other classifiers like Support Vector Machine, Random Forest, solo CNN, and standalone LSTM have been observed at 82%, 85%, 87%, and 86%, respectively. The average accuracy, recall, and F1-score, presented in the table below, shows how well each serve type was predicted. 92, 0. 92, and 0. 92, respectively. The AUC of all serve kinds was more than 0. 90, indicating the discriminating ability of the test in this regard was quite high. High accuracy and a low time complexity were achieved, with the training and validation curves showing a steadily increasing training accuracy of 92% and a validation accuracy that peaked at 91% by epochs 50. The training loss of the model reduces to 0. 20, and the validation loss is still at 0. 30, meaning that despite the high variance, the model is sufficiently generalized and does not overfit.